An adaptive early warning and intervention system for chronic obstructive pulmonary disease

Abstract

This invention discloses an adaptive early warning and intervention system for chronic obstructive pulmonary disease (COPD), relating to the fields of telemedicine and intelligent chronic disease management. The system includes: collecting multi-source physiological and clinical data from COPD patients to obtain a multi-dimensional input dataset; extracting risk feature vectors; constructing and dynamically updating individual health profiles; inputting the risk feature vectors and clinical baseline data into an LSTM dynamic risk prediction model to output risk probabilities; performing adaptive early warning judgments based on risk probabilities and individual health profiles to output early warning levels; dynamically assembling and pushing personalized intervention plans from a structured intervention knowledge base; and continuously updating the adaptive rule base and prediction model parameters by collecting feedback data. This invention solves the technical problems of existing COPD remote management systems, such as frequent false alarms and missed alarms due to static judgment using fixed thresholds, insufficient targeting of early warning and intervention content, long response chains, and a lack of adaptive adjustment capabilities, thereby reducing the occurrence of false alarms and missed alarms.

Description

Technical Field

[0001] This invention relates to the field of telemedicine and intelligent management of chronic diseases, and more specifically, to an adaptive early warning and intervention system for chronic obstructive pulmonary disease. Background Technology

[0002] Chronic obstructive pulmonary disease (COPD) is a chronic respiratory disease characterized by persistent airflow limitation, requiring long-term outpatient monitoring and management. With the development of smart wearable devices and telemedicine technology, COPD remote management systems based on continuous physiological data acquisition are gradually being applied. Existing systems typically collect physiological indicators such as blood oxygen saturation and respiratory rate, combine them with fixed threshold rules to judge the patient's condition, and issue alerts to medical staff or patients when abnormalities occur.

[0003] However, existing COPD remote management systems have the following shortcomings: First, the system uses fixed thresholds to statically judge physiological indicators, failing to distinguish between normal physiological fluctuations and pathological deterioration under different activity states such as rest, exercise, and sleep, leading to frequent false alarms and missed alarms. Second, the system lacks the ability to comprehensively utilize patients' historical health data, history of acute exacerbations, medication regimens, comorbidities, and rehabilitation compliance; the warning and intervention content is generic and lacks specificity. Third, the response chain after a warning is triggered, relying on manual analysis and contact with the patient, is long, easily missing the early intervention window when the condition deteriorates rapidly. Fourth, intervention suggestions are out of touch with patients' actual life scenarios, lacking immediacy and feasibility, resulting in low patient compliance. Fifth, system parameters remain unchanged for a long time after being set, lacking the ability to adaptively adjust based on actual feedback data. These problems prevent existing technology from forming an intelligent closed-loop management capability from continuous monitoring to accurate warnings to personalized intervention. Summary of the Invention

[0004] This invention provides an adaptive early warning and intervention system for chronic obstructive pulmonary disease (COPD), which solves the technical problems in related technologies such as the lack of an individualized dynamic risk prediction and adaptive intervention mechanism for COPD patients, the high false alarm rate caused by rigid early warning rules and the inability to adaptively adjust according to individual patient differences, and the lack of personalized customization capability of intervention programs.

[0005] This invention discloses an adaptive early warning and intervention system for chronic obstructive pulmonary disease, comprising the following steps: Collect patients' real-time physiological signals and clinical baseline data, and integrate them to generate a multi-dimensional input dataset; Preprocessing and feature extraction are performed on the physiological time-series data in the multi-dimensional input dataset to obtain a risk feature vector containing the frequency of daytime blood oxygen decline events, nighttime hypoxia load, respiratory muscle fatigue index, exercise endurance decay slope, and changes in rehabilitation training completion rate. Individual health profiles of patients are constructed based on multi-dimensional input datasets. The individual health profiles include static information fields, historical medical record fields, baseline physiological parameter fields, and lifestyle preference fields. Input the risk feature vector and clinical baseline data into the dynamic risk prediction model to output the probability of acute exacerbation and the probability of disability. Based on the probability of acute exacerbation risk, probability of disability risk, individual health profile, real-time physiological signals and clinical baseline data, the system performs early warning judgments in collaboration with real-time threshold judgments and adaptive rule base matching, and outputs the early warning level. Based on the warning level and individual health profile, intervention content is dynamically assembled from the structured intervention knowledge base to generate personalized intervention plans and push them to patients and medical staff. Collect feedback data after early warning, and update the trigger threshold of the adaptive rule base and the parameters of the dynamic risk prediction model based on the feedback data.

[0006] Furthermore, the method for obtaining the nocturnal hypoxic load is as follows: traverse all sampling points during sleep, determine whether the blood oxygen saturation value of each sampling point is lower than 90%, multiply the number of sampling points that meet the condition by the sampling time interval to obtain the cumulative hypoxic time, and then divide the cumulative hypoxic time by the total sleep duration and convert it into a percentage value to obtain the nocturnal hypoxic load; wherein, the sampling time interval and the total sleep duration use the same time unit.

[0007] Furthermore, the respiratory muscle fatigue index is obtained as follows: the difference between the average amplitude of the respiratory waveform at baseline and the average amplitude of the respiratory waveform in the current monitoring window is divided by the average amplitude of the respiratory waveform at baseline to obtain the relative amplitude attenuation ratio; the standard deviation of the respiratory frequency sequence in the current monitoring window is divided by the mean to obtain the respiratory frequency variation coefficient; the relative amplitude attenuation ratio and the respiratory frequency variation coefficient are multiplied by their respective preset weighting coefficients and then summed to obtain the respiratory muscle fatigue index; wherein, the sum of the two preset weighting coefficients is 1 and both are positive numbers.

[0008] Furthermore, the baseline physiological parameter field is obtained as follows: the physiological time series data during the initial monitoring period are divided into three groups according to the activity intensity label: resting, light activity, and sleep. The mean and standard deviation of each group are calculated separately, and the standard deviation of the mean plus or minus a preset multiple is used as the baseline range for the corresponding activity state. After each acquisition of new physiological time series data and clinical data, the individual health profile recalculates the statistics of the baseline physiological parameter field at a preset update cycle, and writes the new symptom self-assessment records and rehabilitation training completion data into the lifestyle preference field.

[0009] Furthermore, the dynamic risk prediction model employs a long short-term memory network. The input layer of the long short-term memory network receives two data streams: the first stream is a sequence of hourly risk feature vectors from the past several consecutive days as a time-series input, and the second stream is a static input vector of encoded clinical baseline data. Categorical variables are encoded using one-hot encoding, while numerical variables are normalized to the mean based on the range. The long short-term memory network progressively updates the hidden state of the time-series input. The output of the hidden state at the final time step is concatenated with the static input vector and then fed into a fully connected linear mapping layer. After passing through a sigmoid activation function, the probability of acute exacerbation and the probability of disability are output, respectively, both of which are mapped to the interval between zero and one.

[0010] Furthermore, the collaborative execution of early warning judgment through real-time threshold judgment and adaptive rule base matching includes: comparing real-time blood oxygen saturation and respiratory rate values ​​with the baseline range of the corresponding activity state in the individual's health profile; generating a single indicator exceeding the limit when the real-time blood oxygen saturation value is lower than the lower limit of the individual's baseline range and the duration exceeds a preset duration threshold, or when the real-time respiratory rate value is higher than the upper limit of the individual's baseline range and the duration exceeds a preset duration threshold; inputting the single indicator exceeding the limit mark, risk probability, indicator change trend, and indicator composite relationship into the adaptive rule base for matching, and outputting an early warning level based on the matching result; the early warning level is divided into three levels: reminder level, intervention level, and emergency level.

[0011] Furthermore, the trigger thresholds for each rule in the adaptive rule base are dynamically adjusted based on the patient's GOLD grade, history of acute exacerbations, and current risk probability of acute exacerbations. The adjustment method is as follows: starting from the baseline trigger threshold stored in the adaptive rule base, a correction offset is applied to the baseline trigger threshold based on three factors: GOLD grade, number of acute exacerbations in the past year, and current risk probability of acute exacerbations. The three correction offsets are then added together to obtain the personalized trigger threshold currently applicable to the patient. The higher the GOLD grade, the more acute exacerbations, or the greater the risk probability of acute exacerbations, the greater the magnitude of the correction offset that shifts the personalized trigger threshold in a more sensitive direction.

[0012] Furthermore, the dynamic assembly of intervention content from the structured intervention knowledge base includes: filtering the scope of intervention content templates according to the warning level, where the intervention content corresponding to the reminder level warning is pushed to the patient end, the intervention content corresponding to the intervention level warning is pushed to both the patient end and the medical staff end, and the emergency level warning triggers the highest priority alarm and pushes the emergency medical reminder to the patient end and the emergency contact instruction to the medical staff end respectively; according to the medication plan, activity level and previous intervention response records in the individual health profile, the parameterized fields in the selected intervention content template are filled and adjusted to obtain the personalized intervention plan text; the personalized intervention plan text and the associated monitoring data summary are encapsulated into push messages to the patient end and push messages to the medical staff end respectively, and pushed synchronously to the corresponding terminals through the message channel.

[0013] Furthermore, updating the trigger thresholds of the adaptive rule base and the parameters of the dynamic risk prediction model based on feedback data includes: performing correlation analysis between the patient symptom change data and subsequent physiological time-series monitoring data in the feedback data and the risk probability value and warning level at the time of warning triggering, and determining the validity label of this warning. The validity label includes valid warning and over-warning; for rules associated with over-warning, increasing the trigger threshold of the corresponding rule in the adaptive rule base or extending the duration requirement; for rules associated with missed reporting, decreasing the trigger threshold of the corresponding rule or shortening the duration requirement; supplementing the physiological time-series monitoring data and corresponding event labels in the feedback data into the training dataset of the dynamic risk prediction model, and incrementally training the dynamic risk prediction model with a preset retraining cycle to update the model's network parameters; wherein, the threshold change amount adjusted each time does not exceed a preset proportion of the current threshold.

[0014] This invention discloses an adaptive early warning and intervention system for chronic obstructive pulmonary disease, comprising: The data acquisition module is used to collect patients' real-time physiological signals and clinical baseline data, and integrate them to generate a multi-dimensional input dataset; The feature extraction module is used to preprocess and extract features from the physiological time-series data in the multi-dimensional input dataset to obtain risk feature vectors; The health profile module is used to build and dynamically update individual health profiles of patients based on multi-dimensional input datasets. The risk prediction module is used to input risk feature vectors and clinical baseline data into a dynamic risk prediction model and output the probability of acute exacerbation and the probability of disability. The early warning judgment module is used to perform early warning judgments and output the early warning level based on the probability of acute exacerbation risk, the probability of disability risk, individual health profile, real-time physiological signals and clinical baseline data, through real-time threshold judgment and adaptive rule base matching. The intervention plan generation module is used to dynamically assemble intervention content from a structured intervention knowledge base based on the warning level and individual health profile, generate personalized intervention plans, and push them to the patient and medical staff. The feedback update module is used to collect feedback data after an early warning, and update the trigger threshold of the adaptive rule base and the parameters of the dynamic risk prediction model based on the feedback data.

[0015] This invention provides an adaptive early warning and intervention system for chronic obstructive pulmonary disease (COPD). It addresses the technical problems of existing COPD remote management systems, such as frequent false alarms and missed alarms due to static judgment using fixed thresholds, insufficient targeting of early warning and intervention content, long response chains, and a lack of adaptive adjustment capabilities. The system achieves the following technical effects: By constructing an individual health profile and dynamically correcting the early warning threshold based on the patient's baseline level, the early warning judgment can distinguish between normal physiological fluctuations and pathological deterioration signals under different activity states, reducing false alarms and missed alarms; by employing an early warning method that combines real-time threshold judgment with an adaptive rule base, and comprehensively considering multi-dimensional criteria, the early warning results can more accurately reflect the actual evolution of the disease; by integrating risk identification, early warning judgment, and intervention content generation into an automatically executed process, the response time chain from the occurrence of an anomaly to intervention implementation is shortened; by dynamically assembling intervention content based on the early warning level, the patient's current activity state, and the individual health profile, the intervention suggestions have stronger targeting and scenario adaptability; by feeding back the feedback data after the early warning to the adaptive rule base and dynamic risk prediction model for continuous adjustment and incremental training, the system possesses an adaptive iterative capability that continuously updates with changes in patient status and data accumulation. Attached Figure Description

[0016] Figure 1 This is a flowchart of the adaptive early warning and personalized intervention method for COPD patients provided in the embodiments of the present invention; Figure 2 This is a schematic diagram of hourly SpO2 and respiratory rate time-series monitoring provided in an embodiment of the present invention on April 15th; Figure 3 This is a schematic diagram of the Z-score standardized values ​​of each component of the risk feature vector provided in the embodiments of the present invention; Figure 4 This is a schematic diagram showing the daily peak activity intensity and the trend of exercise endurance decline from April 9th ​​to April 15th, provided by an embodiment of the present invention. Figure 5 This is a schematic diagram illustrating the trend of changes in the completion rate and compliance of rehabilitation training over 7 consecutive days, provided in an embodiment of the present invention. Figure 6This is a schematic diagram of the hourly distribution of SpO2 during nighttime sleep (from 22:00 on April 14 to 06:00 on April 15) provided in an embodiment of the present invention; Figure 7 This is a schematic diagram comparing key indicators and thresholds for early warning judgment provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the SpO2 recovery process before and after the early warning (April 15 to April 17) provided in an embodiment of the present invention. Detailed Implementation

[0017] In the fields of telemedicine and chronic disease management, long-term outpatient monitoring and early warning of disease exacerbation risk for patients with chronic obstructive pulmonary disease (COPD) are crucial aspects of continuous management. Currently, COPD remote management systems typically use fixed thresholds to measure blood oxygen saturation (…). Static assessments based on single indicators such as respiratory rate fail to differentiate between normal physiological fluctuations and pathological deterioration under different activity states, including rest, exercise, and sleep, leading to frequent false alarms and missed alarms. Furthermore, existing systems lack the ability to comprehensively utilize patients' historical health data, acute exacerbation history, medication regimens, comorbidities, and rehabilitation adherence. Warnings and interventions are characterized by generalization and standardization, lacking specificity. Simultaneously, the response chain after a warning is triggered, relying on manual analysis and contact with the patient, is lengthy, easily missing the early intervention window for rapid deterioration. Intervention recommendations are disconnected from patients' actual life scenarios, lacking immediacy and feasibility, resulting in low patient adherence and limited overall management effectiveness. The core contradiction of these problems lies in the fact that existing technologies have failed to form an intelligent closed-loop management capability from continuous monitoring to precise warnings to personalized interventions.

[0018] Please see Figure 1 , Figure 2 and Figure 3 According to an embodiment of this invention, an adaptive early warning and intervention system for chronic obstructive pulmonary disease (COPD) is provided to address the aforementioned technical problems. It should be understood that the adaptive early warning and personalized intervention method for COPD patients operates on a computing platform including a cloud server or a local server. The patient's end collects and uploads data through smart wearable devices (including medical-grade smart bracelets, ring-type pulse oximeters, patch-type monitoring devices, or chest straps, etc.) and mobile terminals. The healthcare management end receives early warning information and manages access points through terminal devices.

[0019] The steps of this implementation method are as follows:

[0020] Step 1: Collect multi-source physiological and clinical data from COPD patients to obtain a multi-dimensional input dataset. Through smart wearable devices worn by patients, real-time physiological and behavioral signals are continuously collected, including respiratory waveforms, Physiological time-series data, including respiratory rate, heart rate, activity intensity, and sleep status, are collected. Simultaneously, clinical baseline data is obtained through the patient-side application and the hospital's electronic medical record system interface, including age, gender, GOLD classification, history of acute exacerbations, hospitalization history, medication regimen, comorbidities, smoking history, and symptom self-report information. The physiological time-series data and clinical baseline data are integrated to generate a multi-dimensional input dataset.

[0021] It should be noted that the activity intensity mentioned above refers to the patient's exercise load level as represented by metabolic equivalents (METs), calculated using an accelerometer built into the wearable device. The sleep state mentioned above is a marker of the patient's current sleep / wake stage based on body movement signals and heart rate variability.

[0022] It should be noted that the above physiological time series data were collected hourly, with a continuous collection period of no less than 7 days, to ensure the integrity of the time series required for subsequent feature extraction.

[0023] In this embodiment, to ensure the integrity and consistency of the input data, missing value detection and imputation are performed on the collected physiological time-series data before generating the multi-dimensional input dataset. Specifically, for missing segments in continuous time-series data caused by device detachment or transmission interruption, a linear interpolation algorithm is used to fill in the missing values ​​and obtain a complete continuous time-series record. Data segments with missing durations exceeding a preset duration threshold are marked as invalid segments and excluded from subsequent feature calculations.

[0024] Step 2: Preprocess and extract features from the physiological time-series data in the multi-dimensional input dataset to obtain the risk feature vector. The physiological time-series data in the multi-dimensional input dataset are filtered and denoised to remove motion artifacts and environmental interference, resulting in standardized time-series data. Based on the standardized time-series data, multi-dimensional features for risk identification are extracted to obtain a risk feature vector.

[0025] The aforementioned risk feature vector includes at least the following feature components:

[0026] Sub-step 201: Statistically standardize time series data for daytime Get the number of events with a value lower than the preset daytime threshold. Decreasing event frequency .

[0027] Furthermore, the aforementioned preset daytime threshold is based on the patient's individual health profile at rest. The judgment boundary, determined by combining the lower limit of the baseline range and the clinically accepted reference value, is used to distinguish between normal daily physiological fluctuations and pathological events of decreased oxygen saturation.

[0028] Sub-step 202: Calculate sleep duration A cumulative time of less than 90% of total sleep time indicates a nighttime low oxygen load. Low oxygen load at night The calculation expression is: ; in, This represents the total number of sampling points during sleep. The sampling point number is... The range of values ​​is from arrive positive integers, For the first Blood oxygen saturation values ​​at each sampling point The sampling time interval, Total sleep duration This is an indicator function; it takes the value 1 if the condition within the parentheses is true, and 0 otherwise. (The above...) and Using the same time unit makes the nighttime low oxygen load This is a dimensionless percentage value.

[0029] Sub-step 203: Calculate the respiratory muscle fatigue index based on the amplitude attenuation characteristics of the respiratory waveform signal and the coefficient of variation of the respiratory rate. Respiratory muscle fatigue index The calculation expression is: ; in, The baseline respiratory waveform average amplitude. This represents the average amplitude of the respiratory waveform within the current monitoring window. This represents the coefficient of variation of respiratory rate within the current monitoring window. and The weighting coefficients are preset and satisfy the following conditions: , and All are positive numbers. The above This is the relative attenuation ratio of the amplitude. This is the ratio of the standard deviation of respiratory rate to the mean. Both are dimensionless ratios with consistent dimensions, and can be directly weighted and summed.

[0030] Among them, the coefficient of variation of respiratory rate Defined as the ratio of the standard deviation to the mean of the respiratory rate sequence within the current monitoring window, i.e. , The standard deviation of respiratory rate within the window. This represents the average respiratory rate within the window.

[0031] Furthermore, the aforementioned respiratory muscle fatigue index The physical meaning is: when a patient's respiratory muscles are fatigued, the amplitude of the respiratory waveform decreases relative to the baseline period, and at the same time, the regularity of the respiratory rhythm decreases, leading to an increase in the coefficient of variation of the respiratory rate. The weighted sum of these two factors results in the respiratory muscle fatigue index. An increase in the numerical value reflects a greater degree of respiratory muscle fatigue. When equal and When at a normal level, the respiratory muscle fatigue index A value closer to a lower value indicates that the respiratory muscles are in good condition.

[0032] Sub-step 204: Perform linear regression fitting on the peak activity intensity data over multiple consecutive days to obtain the slope of exercise endurance decay. Slope of exercise endurance decline Characterizes the trend of changes in patients' exercise tolerance over time.

[0033] Please see Figure 4 and Figure 5 Furthermore, the aforementioned slope of exercise endurance decay The slope of the linear regression line is obtained by fitting time (days) as the independent variable and peak daily activity intensity (METs) as the dependent variable. (Slope of exercise endurance decay) A negative value indicates that the patient's exercise tolerance decreases over time, and the slope of the exercise tolerance decline is... The larger the absolute value, the more significant the downward trend; the slope of exercise endurance decay. A positive value or close to zero indicates that exercise endurance remains stable or has improved.

[0034] Sub-step 205: Calculate the rate of change of rehabilitation training completion data within the continuous monitoring period, and obtain the change in rehabilitation training completion rate. Specifically, the change in rehabilitation training completion rate. Defined as the difference between the rehabilitation training completion rate at the end of the current monitoring period and the rehabilitation training completion rate at the beginning of the monitoring period, i.e. ,in To monitor the completion rate of rehabilitation training at the end of the cycle, To monitor the rehabilitation training completion rate at the beginning of the cycle, the values ​​are all within the range of .

[0035] The above feature components are combined to generate a risk feature vector. .

[0036] It should be noted that the physical meanings and dimensions of the above feature components are different. Before combining them into a risk feature vector, each feature component is subjected to Z-score standardization to transform each component into a dimensionless value with a mean of 0 and a standard deviation of 1, thereby eliminating the influence of dimensional differences on subsequent model inputs and obtaining a standardized risk feature vector.

[0037] It should be noted that the above respiratory muscle fatigue index The mean amplitude of the baseline respiratory waveform This refers to the statistical mean of the amplitude of the patient's respiratory waveform at rest during the initial monitoring period, and the baseline average amplitude of the respiratory waveform. In step 3, these parameters are stored in the individual health profile as part of the baseline physiological parameters.

[0038] Please see Figure 6 , Figure 7 and Figure 8 In this embodiment, to extract more stable temporal feature representations, the standardized time-series data is segmented using a sliding window before calculating each feature component. Specifically, the continuous time-series data is segmented with a preset window length and step size. The statistics of each feature are calculated within each window, and the statistics of multiple windows are aggregated to obtain smoothed feature component values, thereby reducing the impact of short-term abnormal fluctuations on the feature extraction results.

[0039] Step 3: Based on the multi-dimensional input dataset, construct and dynamically update the individual health profile of the patient.

[0040] The system reads clinical baseline data and physiological time-series data from a multi-dimensional input dataset to generate an individual health profile for each patient. The individual health profile includes the following four data fields: The first category consists of static information fields, including the patient's age, gender, GOLD classification, and smoking history.

[0041] The second category is historical medical record fields, including the number of previous acute exacerbations, hospitalization history, current medication regimen and dosage.

[0042] The third category consists of baseline physiological parameter fields, obtained through statistical analysis of standardized time-series data during the initial monitoring period, including data from patients at rest, during mild activity, and during sleep. Baseline range, respiratory rate baseline range, and heart rate baseline range.

[0043] The fourth category is the lifestyle preference field, which includes daily activity levels, rehabilitation exercise habits, symptom self-assessment records, and compliance records.

[0044] By associating and storing the above four types of data fields, a structured individual health profile data record is generated.

[0045] It should be noted that the baseline ranges for each state in the aforementioned baseline physiological parameter fields refer to the baseline ranges for the corresponding state during the initial monitoring period. This is achieved by grouping the physiological time-series data according to activity intensity labels (resting, light activity, sleep), calculating the mean and standard deviation for each group, and then using the standard deviation (mean plus or minus a preset multiple) as the baseline range. For example, the baseline range for the resting state... Baseline range is ,in In a resting state The mean, For the corresponding standard deviation, This is a preset multiple.

[0046] In this embodiment, to ensure that the individual health profile reflects the patient's latest health status, the baseline physiological parameter field and lifestyle preference field in the individual health profile are periodically updated after each acquisition of new physiological time-series data and clinical data. Specifically, the statistics of the baseline range are recalculated at a preset update cycle, and new symptom self-assessment records and rehabilitation training completion data are written into the lifestyle preference field to obtain the updated individual health profile.

[0047] Step 4: Input the risk feature vector and clinical baseline data into the dynamic risk prediction model, and output the risk probability.

[0048] Risk feature vector As a time-series input, clinical baseline data from the multi-dimensional input dataset is encoded into a static input vector, which is then fed into the dynamic risk prediction model to output the patient's risk probability value within a preset prediction time window. This risk probability value includes the probability of acute exacerbation. and the probability of disability .

[0049] The aforementioned dynamic risk prediction model employs a Long Short-Term Memory (LSTM) network. The LSTM input layer receives two data streams: the first is a sequence of hourly risk feature vectors from the past seven consecutive days, constituting the temporal input; the second is a static input vector of encoded clinical baseline data. Categorical variables (such as GOLD classification and comorbidity type) are encoded using one-hot encoding, while numerical variables (such as age and number of acute exacerbations) are normalized using mean normalization based on the range. The encoded dimensions are concatenated to form the static input vector. The LSTM progressively updates the hidden state of the temporal input, extracting temporal risk features. The output of the hidden state at the final time step is concatenated with the static input vector and then fed into a fully connected linear mapping layer, outputting the probability of acute exacerbation risk. and the probability of disability Both are mapped to the Sigmoid activation function. Interval.

[0050] Furthermore, the output dimension of the aforementioned fully connected linear mapping layer is 2, corresponding to the probability of acute exacerbation risk. and the probability of disability Two risk probability values. The Sigmoid activation function independently maps the real-valued output of the fully connected linear mapping layer to... The range makes the probability of acute exacerbation... and the probability of disability All of these can be interpreted as probability estimates of the occurrence of the corresponding risk event within a preset prediction time window.

[0051] Furthermore, the aforementioned preset prediction time window is a fixed duration interval extending into the future from the current time point, specifically taking the value of 3 months or 6 months, used to define the probability of acute exacerbation risk. and the probability of disability The corresponding target time range for prediction is the probability of acute exacerbation output by the dynamic risk prediction model. This represents an estimate of the probability of a patient experiencing an acute exacerbation within the next 3 or 6 months, indicating a risk of disability. This represents an estimate of the probability that a patient will experience a disabling event within the same time window.

[0052] Long Short-Term Memory (LSTM) networks are trained in a supervised manner using labeled data from a historical patient population during the training phase. To provide a binary label for whether an acute exacerbation or disability event will occur within the prediction time window (3 months or 6 months), the loss function is binary cross-entropy loss, and the optimization strategy is Adam optimization algorithm.

[0053] In this embodiment, to improve the stability of risk prediction, data augmentation is performed on the input sequence before inputting the risk feature vector sequence into the Long Short-Term Memory (LSTM) network. Specifically, short-term missing segments in the hourly data of 7 consecutive days are filled in using forward padding, and each feature dimension is normalized using Z-score to obtain a normalized input sequence, which is then input into the LSTM network for risk probability calculation.

[0054] In this embodiment, based on step 4, the risk probability value output by the Long Short-Term Memory (LSTM) network is further marked as an abnormal state. Specifically, when the risk probability of acute exacerbation... or probability of disability When the preset risk threshold is exceeded, the current time point is marked as an abnormal state, and the abnormal state mark along with the risk probability value is transmitted to the subsequent early warning judgment process.

[0055] Furthermore, the aforementioned preset risk threshold is a classification decision boundary value determined based on the receiver operating characteristic curve (ROC curve) on the validation set and using the maximization of the Youden exponent as the criterion after the dynamic risk prediction model training is completed. The preset risk threshold is stored in the configuration parameters after the system is deployed and can be recalibrated based on the new validation set results after the incremental training of the dynamic risk prediction model is completed in step 7.

[0056] Step 5: Based on risk probability, individual health profile, and real-time physiological data, perform adaptive early warning judgment and output the early warning level.

[0057] It reads the risk probability value, the baseline physiological parameters in the individual health profile, and the currently collected real-time physiological time series data, comprehensively performs early warning judgment, and outputs the early warning level.

[0058] The early warning judgment process involves the coordinated work of two sub-processes: real-time threshold judgment and adaptive rule base matching. Sub-step 501: Perform threshold judgment on the current real-time physiological time series data. The values ​​of respiratory rate and respiratory rate are compared with the baseline range of the corresponding activity state in the individual's health profile. When real-time... When the value is lower than the lower limit of the individual baseline range and the duration exceeds the preset duration threshold, or when the real-time respiratory rate value is higher than the upper limit of the individual baseline range and the duration exceeds the preset duration threshold, a single indicator exceeding the limit flag is generated.

[0059] Sub-step 502: Input the single indicator exceeding the limit marker, risk probability value, indicator change trend, and indicator composite relationship into the adaptive rule base for matching. The adaptive rule base stores multiple warning trigger rules, each containing trigger conditions, a corresponding warning level, and applicable patient risk stratification. The trigger conditions in the adaptive rule base include not only single indicator thresholds but also the direction of indicator change trends, multi-indicator combination conditions, and duration requirements. Based on the matching results, output the warning level.

[0060] The aforementioned warning levels are divided into three levels: Level 1 is the alert level, triggered when a single indicator deviates slightly from the baseline range or the risk probability value is at a moderate level; Level 2 is the intervention level, triggered when multiple indicators deviate from the baseline range simultaneously or the risk probability value exceeds a high threshold; Level 3 is the emergency level, triggered when... Triggered when serious abnormalities occur, such as a respiratory rate below 88% for more than 5 minutes or a respiratory rate exceeding 30 breaths / min.

[0061] It should be noted that the trigger thresholds for each rule in the aforementioned adaptive rule base are not fixed values, but rather dynamically adjusted based on the patient's individual health profile, including GOLD classification, history of acute exacerbations, and current risk probability. For example, for patients with frequent history of acute exacerbations or high GOLD classifications, the adaptive rule base automatically employs more sensitive triggering conditions, namely, lowering the exceedance thresholds for each indicator and shortening the duration requirements.

[0062] Furthermore, the specific method of the above dynamic correction is as follows: starting from the baseline trigger threshold stored in the adaptive rule base, based on the patient's GOLD classification, the number of acute exacerbations in the past year, and the current probability of acute exacerbation risk. Each of the three factors applies a correction offset to the baseline trigger threshold. The sum of these three correction offsets yields the personalized trigger threshold applicable to the patient. Higher GOLD grades, more frequent acute exacerbations, or a higher risk of acute exacerbations are associated with these factors. The larger the value, the greater the magnitude of the shift in the personalized trigger threshold towards a more sensitive direction, thereby enabling early warning for high-risk patients.

[0063] In this embodiment of the application, in order to reduce false alarms caused by transient physiological fluctuations, trend confirmation processing is also performed on the real-time physiological data in sub-step 501. Specifically, when a single indicator exceeds the limit for the first time, a single indicator exceeding the limit marker is not immediately generated. Instead, continuous sampling is performed within a preset observation window, and the deviation between the mean value of the indicator within the observation window and the baseline range is calculated. Only when the deviation exceeds a preset deviation threshold is a single indicator exceeding the limit marker generated, and the exceeding judgment result after trend confirmation is obtained.

[0064] In this embodiment of the application, based on step 5, a composite verification of the warning level is further performed. Specifically, the warning level output by the adaptive rule base is cross-validated with the risk probability value output by the Long Short-Term Memory (LSTM) network. When there is an inconsistency between the warning level determined by the adaptive rule base and the risk level indicated by the risk probability value, the higher warning level is taken as the final output to reduce the possibility of missed detections.

[0065] Step 6: Based on the warning level and individual health profile, dynamically assemble intervention content from the structured intervention knowledge base, generate personalized intervention plans, and push them to both patients and healthcare professionals.

[0066] The system reads the warning level, the specific rule identifier that triggers the warning level, the patient's current activity status label, real-time physiological parameter values, and individual health profile. It then retrieves intervention content items that match the above input conditions from the structured intervention knowledge base and dynamically assembles them to generate a personalized intervention plan.

[0067] The structured intervention knowledge base described above stores multiple types of intervention content templates. Each type of template is associated with an applicable warning level, an applicable activity status, and an applicable range of patient characteristics. The intervention content templates include at least breathing training guidance, rest and activity suggestions, medication reminders, oxygen therapy suggestions, a symptom self-assessment checklist, and emergency medical attention reminders.

[0068] The assembly process of a personalized intervention plan is as follows: Sub-step 601: Filter the scope of intervention content templates based on the warning level. Level 1 warnings correspond to intervention content pushed to the patient end, including breathing training guidance and activity adjustment suggestions; Level 2 warnings correspond to intervention content pushed to both the patient and healthcare staff ends, with the patient receiving oxygen therapy adjustment suggestions and medication inhalation reminders, and the healthcare staff receiving details of abnormal events and trend curve data; Level 3 warnings trigger the highest priority alarm, with the patient receiving an emergency medical attention reminder, and the healthcare staff receiving emergency contact instructions and a complete monitoring data summary.

[0069] Sub-step 602: Based on the medication regimen, activity level, and previous intervention response records in the individual health profile, fill in and adjust the parameterized fields in the selected intervention content template to obtain the personalized intervention plan text.

[0070] Sub-step 603: Encapsulate the personalized intervention plan text and the associated monitoring data summary into push messages for the patient and healthcare staff respectively, and push them synchronously to the corresponding terminals through the message channel.

[0071] It should be noted that the intervention content templates in the aforementioned structured intervention knowledge base are standardized entries pre-reviewed and entered by clinical experts. Each intervention content template entry includes an intervention type label, a description of applicable conditions, parameterized content text, and an execution priority field. The parameterized content text contains replaceable placeholders, which are automatically filled in by the system based on specific data values ​​in the patient's individual health profile during assembly. The final result is a personalized intervention plan text containing specific values ​​and actionable instructions, which is directly pushed to the patient's and healthcare provider's terminal devices.

[0072] In this embodiment, to improve the adaptability of the intervention content to the patient's current scenario, sub-step 602 further performs scenario adaptation processing on the intervention content based on the patient's current activity status label. Specifically, when the patient is asleep, intervention content template items that do not require immediate active operation are selected first (such as delaying the push of medication reminders until the patient wakes up); when the patient is active, intervention content template items related to activity regulation are selected first (such as suggestions to reduce activity intensity), thus obtaining a personalized intervention plan after scenario adaptation.

[0073] Step 7: Collect feedback data after the early warning, and update the parameters of the adaptive rule base and dynamic risk prediction model based on the feedback data.

[0074] After each warning is triggered and a personalized intervention plan is pushed out, subsequent feedback data is collected, including records of patients' response to the personalized intervention plan, records of medical staff's handling operations, data on changes in patients' symptoms within a preset time period after the warning, and subsequent physiological time-series monitoring data.

[0075] Sub-step 701: Correlate the patient symptom change data and subsequent physiological time-series monitoring data in the feedback data with the risk probability value and warning level at the time of warning triggering to determine the validity label of this warning. When the patient's symptoms do indeed worsen or acute exacerbate after the warning is triggered, it is marked as a valid warning; when the patient's physiological indicators spontaneously return to the baseline range within a short period of time after the warning is triggered and no symptom changes occur, it is marked as an over-warning.

[0076] Sub-step 702: Based on the validity label and the associated trigger rule identifier, adjust the trigger threshold and weight parameters of the corresponding rules in the adaptive rule base. For rules associated with excessive warnings, appropriately increase the trigger threshold or extend the duration requirement of the corresponding rules in the adaptive rule base; for rules associated with missed warnings, appropriately decrease the trigger threshold or shorten the duration requirement of the corresponding rules in the adaptive rule base.

[0077] Sub-step 703: Supplement the physiological time-series monitoring data and corresponding event labels in the feedback data into the training dataset of the Long Short-Term Memory Network (LSTM), perform incremental training on the Long Short-Term Memory Network (LSTM) with a preset retraining cycle, update the network parameters of the Long Short-Term Memory Network (LSTM), and obtain the optimized dynamic risk prediction model.

[0078] It should be noted that in the above-mentioned determination of the validity of the label, the time window for determining whether the warning is effective is the preset observation period after the warning is triggered (e.g., 48 to 72 hours). Within the preset observation period, a comprehensive assessment is made of whether the patient's symptoms have worsened, whether unplanned medical treatment or medication adjustments have been made.

[0079] In this embodiment of the application, to avoid unstable warning behavior due to excessive adjustments to the adaptive rule base, an upper limit constraint is set on the single adjustment range of the rule threshold in sub-step 702. Specifically, the threshold change in each adjustment does not exceed a preset proportion of the current threshold, and cumulative adjustment is only performed when the adjustment direction of the same rule remains consistent across multiple consecutive update cycles, thereby obtaining a stably updated adaptive rule base.

[0080] In this embodiment of the application, based on step 7, the patient response behavior records and personalized intervention plan implementation status in the feedback data are also written back to the lifestyle preference field and compliance record of the individual health profile, so that the individual health profile continuously reflects the patient's acceptance and implementation of different types of intervention measures, thereby providing updated personalized basis for the assembly of intervention content in the subsequent step 6.

[0081] This embodiment provides an adaptive early warning and intervention system for chronic obstructive pulmonary disease. In step 3, an individual health profile is constructed for each patient, including static information, historical medical records, baseline physiological parameters, and lifestyle preferences. In step 5, the early warning threshold is dynamically adjusted based on the individual health profile. This makes the early warning judgment no longer dependent on a uniform fixed threshold, but rather differentiated based on the patient's baseline level, disease stage, and current activity status. Therefore, it can more accurately distinguish between normal physiological fluctuations and pathological deterioration signals, reducing the occurrence of false alarms and missed alarms.

[0082] Because step 5 employs a real-time threshold judgment and adaptive rule base collaborative early warning method, which comprehensively considers multiple dimensions of criteria such as single indicator exceeding limits, indicator change trends, multi-indicator composite relationships, and duration, rather than judging only the instantaneous value of a single indicator, the early warning results can more accurately reflect the actual evolution of the disease.

[0083] Because in step 4, a Long Short-Term Memory (LSTM) network is used to predict the risk probability of continuous time series data, and in steps 5 and 6, the warning classification and intervention content assembly are automatically completed based on the risk probability and warning level, the risk identification, warning judgment and intervention content generation are integrated into an automatically executed process, no longer relying on manual analysis and manual contact, thus shortening the response time chain from the occurrence of anomalies to the implementation of intervention.

[0084] Because in step 6, intervention content is dynamically assembled from the structured intervention knowledge base based on the warning level, the patient's current activity status, and individual health profile, the generated personalized intervention plan is parameterized according to the patient's specific medication plan, activity level, and past response records, rather than using a uniform intervention content template. Therefore, the intervention suggestions are more targeted and adaptable to different scenarios, which helps to improve the patient's understanding of the intervention measures and their willingness to implement them.

[0085] Because in step 7, the feedback data after each warning is fed back to the adaptive rule base and dynamic risk prediction model for continuous adjustment and incremental training, the warning logic and risk prediction capabilities can be continuously updated as the patient's condition changes and data accumulates. Therefore, the system has adaptive iteration capabilities, which overcomes the limitation of traditional systems that remain unchanged for a long time after being set once, and forms a closed-loop management capability from monitoring to warning to intervention and then to feedback optimization.

[0086] A COPD patient (patient number PT-20XX-0873) was enrolled in a remote chronic disease management platform for outpatient intelligent monitoring in the spring of 20XX. The patient was a 62-year-old male, GOLD III, with two acute exacerbations within the past year, and concurrent hypertension. He was on long-term use of tiotropium bromide inhaler (18 μg / day) combined with salmeterol / fluticasone combination therapy (50 / 500 μg, twice daily). He had a 38-year smoking history (quit 5 years ago). The patient wore a medical-grade smart bracelet and a ring-type pulse oximeter, uploading data to the cloud server via a patient-side application. The system initiated initial monitoring on April 1, 20XX, and entered the formal monitoring phase after collecting baseline data for 7 consecutive days. The following describes the specific operation of each step using the complete data processing flow on April 15, 20XX (the 15th monitoring day) as an example.

[0087] In step 1, the system continuously collects physiological time-series data of PT-20XX-0873 hourly using the smart bracelet and ring-type pulse oximeter worn by the patient, and simultaneously retrieves clinical baseline data from the hospital's electronic medical record system interface, integrating them to generate a multi-dimensional input dataset. During the missing value detection phase, the system found that the patient's SpO2 and respiratory rate data were missing for one hour between 02:00 and 03:00 on April 15th due to the bracelet falling off. The missing duration did not exceed the preset threshold (4 hours), so a linear interpolation algorithm was used to complete the data for that period. A total of 24 valid time points of physiological data were collected on April 15th. Combined with the clinical baseline data, the multi-dimensional input dataset was completed.

[0088] Table 1. PT-20XX-0873 Multidimensional Input Dataset (Partial Fields):

[0089] In step 2, the system extracts features from the standardized time-series data of April 15, calculates each feature component in turn, and finally combines them into a risk feature vector.

[0090] Sub-step 201: The system counts the number of events where SpO2 was lower than the preset daytime threshold (93%, determined based on the lower limit of the patient's resting baseline and comprehensive clinical reference values) between 06:00 and 22:00 on April 15th (16 time points during the day). A total of 5 events with a decrease were detected. The frequency of daytime SpO2 decrease events is f.desat =5 times.

[0091] Sub-step 202: The system extracts SpO2 data from 22:00 on April 14th to 06:00 on April 15th (sleep period), with a total of 8 sampling points (M=8), a sampling time interval of 60 minutes, and a total sleep duration D. sleep =480 minutes. Among them, the sampling points with SpO2 below 90% are 04:00 (88%, satisfied) and 05:00 (89%, satisfied), a total of 2 sampling points that meet the condition; ; Sub-step 203: Baseline respiratory waveform mean amplitude =0.82 (normalized units), the average amplitude of respiratory waveform during the current monitoring window (00:00 to 24:00 on April 15th). =0.61, the mean respiratory rate within the current window is 21.4 breaths / min, the standard deviation is 4.3 breaths / min, and the coefficient of variation (CV) is 0.61. RR =4.3 / 21.4=0.201, take the weighting coefficients α=0.6, β=0.4: ; Sub-step 204: The system performs linear regression fitting on the daily activity intensity peaks from April 9th ​​to April 15th (7 consecutive days), using the day number as the independent variable and the peak METs as the dependent variable, to obtain the slope k of exercise endurance decay. ex =-0.18 METs / day indicates that the patient's exercise tolerance is showing a continuous downward trend.

[0092] Sub-step 205: At the beginning of the monitoring period (April 9th), the rehabilitation training completion rate was 0.75; at the end of the monitoring period (April 15th), the rehabilitation training completion rate was 0.52. The change in rehabilitation training completion rate ΔC rehab =0.52-0.75=-0.23, indicating a significant decrease in rehabilitation compliance during the period.

[0093] Each feature component is standardized by Z-score and then combined into a standardized risk feature vector.

[0094] Table 2. Components of the Risk Feature Vector of PT-20XX-0873 (April 15th):

[0095] In step 3, during the initial monitoring period (April 1st to April 7th), the system uses standardized time-series data to group and statistically analyze data according to three activity status labels: resting, light activity, and sleep, to construct an individual health profile for PT-20XX-0873. This round of updates is completed on April 15th, with the latest symptom self-assessment records and rehabilitation training completion data written into the lifestyle preference field.

[0096] Example of calculating the baseline range of SpO2 under resting conditions: In the initial monitoring period, the mean SpO2 under resting conditions was 94.2%, the standard deviation was 1.1%, and the preset multiplier κ=2, then: .

[0097] Table 3 Core Fields of Individual Health Profile (PT-20XX-0873, April 15th Version):

[0098] In step 4, the system uses the hourly standardized risk feature vector sequence of PT-20XX-0873 over the past 7 consecutive days (April 9th ​​to April 15th) as time-series input, and encodes its clinical baseline data into a static input vector (GOLD III level one-hot encoding is [0,0,1,0], age normalization value is 0.71, number of acute exacerbations normalization value is 0.67, etc.), and inputs both into the LSTM model. After the LSTM progressively updates the hidden state of the 7-day time-series data, it concatenates the final time-step hidden state with the static input vector, passes it through a fully connected linear mapping layer and a Sigmoid activation function, and outputs two risk probability values.

[0099] The preset prediction time window is the next 3 months, and the model output is: the probability of acute exacerbation P. ae =0.81, probability of disability P dis =0.47. The system will set P ae =0.81 compared to the preset risk threshold of 0.65 (determined based on the Youden index of the validation set ROC curve), P ae If the threshold is exceeded, the current time point is marked as an abnormal state, and the abnormal state mark, along with the risk probability value, is transmitted to the early warning judgment process.

[0100] In step 5, the system reads P ae =0.81, baseline physiological parameters of individual health profile and real-time physiological data on April 15, to perform adaptive early warning judgment.

[0101] Sub-step 501: The system compares the real-time SpO2 value of 91% at 06:00 on April 15th with the lower limit of the baseline range of 92.0% in the patient's resting state. 91% < 92.0%, triggering an over-limit detection. The system continuously samples within a preset observation window (30 minutes). The average SpO2 value from 06:00 to 06:30 is 90.8%, deviating from the lower limit of the baseline by 1.2 percentage points, exceeding the preset deviation threshold (0.8 percentage points), thus generating an SpO2 over-limit marker. Simultaneously, the real-time respiratory rate at 06:00 is 23 breaths / min, higher than the upper limit of the baseline range in the awake state after sleep (19.6 breaths / min), and similarly, after trend confirmation, a respiratory rate over-limit marker is generated.

[0102] Sub-step 502: The system will mark the two items as exceeding the limit, P ae =0.81 (high risk level), a continuous decreasing trend of SpO2, and a combined condition of SpO2 and respiratory rate exceeding limits were used for matching in an adaptive rule base. Due to the patient's GOLD grade III, two acute exacerbations in the past year, and P... ae =0.81, the adaptive rule base has applied three corrective offsets to the patient's baseline trigger threshold, and the personalized trigger threshold is more sensitive than the baseline value. The matching result hits the secondary intervention rule (multiple indicators simultaneously deviate from baseline and the risk probability exceeds a higher threshold), outputting a secondary warning (intervention level). After composite validation, the P-value output by LSTM is... ae =0.81, which is consistent with the level 2 warning risk level, and the final output warning level is confirmed as level 2.

[0103] Table 4 Key data in the judgment process of PT-20XX-0873 early warning (April 15, 06:30):

[0104] In step 6, the system reads the secondary warning level, trigger rule identifier (rule number RUL-II-0047), patient's current activity status (awake), real-time SpO2 (90.8%), respiratory rate (23 breaths / min), and individual health profile, retrieves matching entries from the structured intervention knowledge base, and dynamically assembles a personalized intervention plan.

[0105] Sub-step 601: The intervention content corresponding to the Level 2 early warning is pushed to both the patient and medical staff simultaneously. The scope for the patient is limited to oxygen therapy adjustment suggestions and medication inhalation reminders, while the scope for the medical staff is limited to abnormal event details and trend curve data.

[0106] Sub-step 602: The system fills in parameterized fields based on the individual's health profile: the medication regimen field is filled with "tiotropium bromide 18μg + salmeterol / fluticasone 50 / 500μg". The current activity status is awake, so no delay is needed for push notifications; instant push content is generated directly. Because the patient has a missed inhaled medication record this cycle, the system adds an extra prompt to the medication reminder template: "A missed medication was detected this cycle. Please confirm that today's inhaler has been used on time."

[0107] Sub-step 603: The push messages from the patient's end and the push messages from the medical staff's end are encapsulated and pushed synchronously through the message channel at 06:35 on April 15.

[0108] Table 5 Summary of the Personalized Intervention Program for PT-20XX-0873 (April 15):

[0109] In step 7, the system collects feedback data 48 hours after the warning is triggered (i.e., 06:30 on April 17) and carries out closed-loop updates.

[0110] Sub-step 701: Feedback data shows that the patient visited a community clinic on their own at 09:00 on April 15th. The doctor adjusted the frequency of inhaler use. On April 16th, the average SpO2 recovered to 93.1%, and no hospitalization occurred, but a mild acute exacerbation was confirmed. The system marks this warning as a valid warning, with a validity label of 1.

[0111] Sub-step 702: Effective early warning associated with rule number RUL-II-0047, no need to increase the trigger threshold; at the same time, the system checks whether there are any missed reports during the period from April 8th to April 14th, and finds that SpO2 had a single brief drop on April 12th but did not trigger an early warning. The system shortens the duration requirement for RUL-II-0047 from 30 minutes to 22 minutes (the single adjustment range is 26.7% of the current value, which does not exceed the preset upper limit of 30%), so that the rule is triggered more promptly in similar patients.

[0112] Sub-step 703: Add the physiological time-series monitoring data from April 9th ​​to April 17th and the label (y=1) of this mild acute exacerbation event to the LSTM training dataset, and perform incremental training at the next expiration according to the preset retraining cycle (every 30 days). At the same time, the patient's response behavior record (actively seeking medical attention, following oxygen therapy recommendations) and the implementation status of the intervention plan are written back to the lifestyle preference field of the individual health profile, and the compliance record is updated to "good response to oxygen therapy recommendations, good response to medication reminders".

[0113] Table 6 Feedback Data and Rule Update Record for PT-20XX-0873 (April 17th):

[0114] Statement: The patient information in the above embodiments was obtained through legal means and has been disclosed to the public with the consent of the patients.

[0115] The data flow throughout the entire implementation process demonstrates a clear logical progression: the multi-dimensional input data collected in step 1 (including interpolated and completed physiological time-series data and clinical baseline data) directly provides the raw input for feature extraction in step 2. desat L 90 ,RMFI,k ex ΔC rehab The calculation of all five feature components depends on the time-series data from step 1; the baseline physiological parameters (such as the lower limit of the resting SpO2 baseline range of 92.0%) in the individual health profile constructed in step 3... =0.82) was directly used for feature calculation in step 2 and for real-time threshold comparison in step 5, demonstrating the core supporting role of profile data throughout multiple steps; the standardized risk feature vector sequence output from step 2 and the clinical baseline encoding vector from step 1 were jointly fed into the LSTM model in step 4, and the output P ae =0.81 directly participates in the early warning level judgment and rule threshold dynamic correction in step 5; the secondary early warning level output in step 5 triggers the intervention content assembly in step 6, and the specific values ​​in the intervention plan (90.8%, 23 times / min, medication plan) all come from the data stream of the previous steps; the feedback data collected in step 7 is fed back to the rule base and LSTM training set in the form of validity labels, and is also written back to the individual health profile in the form of compliance records, thereby closing the complete closed loop from monitoring and collection to early warning intervention and then to feedback optimization, so that the execution of steps 3, 5 and 6 in the next monitoring cycle can obtain updated input basis.

[0116] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. An adaptive early warning and intervention method for chronic obstructive pulmonary disease, characterized in that, Includes the following steps: Collect patients' real-time physiological signals and clinical baseline data, and integrate them to generate a multi-dimensional input dataset; Preprocessing and feature extraction are performed on the physiological time-series data in the multi-dimensional input dataset to obtain a risk feature vector containing the frequency of daytime blood oxygen decline events, nighttime hypoxia load, respiratory muscle fatigue index, exercise endurance decay slope, and changes in rehabilitation training completion rate. Individual health profiles of patients are constructed based on multi-dimensional input datasets. The individual health profiles include static information fields, historical medical record fields, baseline physiological parameter fields, and lifestyle preference fields. Input the risk feature vector and clinical baseline data into the dynamic risk prediction model to output the probability of acute exacerbation and the probability of disability. Based on the probability of acute exacerbation risk, probability of disability risk, individual health profile, real-time physiological signals and clinical baseline data, the system performs early warning judgments in collaboration with real-time threshold judgments and adaptive rule base matching, and outputs the early warning level. Based on the warning level and individual health profile, intervention content is dynamically assembled from the structured intervention knowledge base to generate personalized intervention plans and push them to patients and medical staff. Collect feedback data after early warning, and update the trigger threshold of the adaptive rule base and the parameters of the dynamic risk prediction model based on the feedback data.

2. The adaptive early warning and intervention method for chronic obstructive pulmonary disease according to claim 1, characterized in that, The method for obtaining nocturnal hypoxia load is as follows: traverse all sampling points during sleep, determine whether the blood oxygen saturation value of each sampling point is lower than 90%, multiply the number of sampling points with blood oxygen saturation values ​​lower than 90% by the sampling time interval to obtain the cumulative hypoxia time, and then divide the cumulative hypoxia time by the total sleep duration and convert it into a percentage value to obtain the nocturnal hypoxia load; wherein, the sampling time interval and the total sleep duration use the same time unit.

3. The adaptive early warning and intervention method for chronic obstructive pulmonary disease according to claim 1, characterized in that, The respiratory muscle fatigue index is obtained as follows: the difference between the average amplitude of the respiratory waveform at baseline and the average amplitude of the respiratory waveform in the current monitoring window is divided by the average amplitude of the respiratory waveform at baseline to obtain the relative amplitude attenuation ratio; the standard deviation of the respiratory frequency sequence in the current monitoring window is divided by the mean to obtain the respiratory frequency variation coefficient; the relative amplitude attenuation ratio and the respiratory frequency variation coefficient are multiplied by their respective preset weighting coefficients and then summed to obtain the respiratory muscle fatigue index; wherein, the sum of the two preset weighting coefficients is 1 and both are positive numbers.

4. The adaptive early warning and intervention method for chronic obstructive pulmonary disease according to claim 1, characterized in that, The baseline physiological parameter fields were obtained as follows: the physiological time series data during the initial monitoring period were divided into three groups according to the activity intensity label: resting, light activity and sleep. The mean and standard deviation of each group of data were calculated separately, and the standard deviation of the mean plus or minus a preset multiple was used as the baseline range of the corresponding activity state. After acquiring new physiological time-series data and clinical data each time, the individual health profile recalculates the statistics of the baseline physiological parameter fields at a preset update cycle, and writes the new symptom self-assessment records and rehabilitation training completion data into the lifestyle preference field.

5. The adaptive early warning and intervention method for chronic obstructive pulmonary disease according to claim 1, characterized in that, The dynamic risk prediction model uses a long short-term memory network. The input layer of the long short-term memory network receives two types of data: the first type is the hourly risk feature vector sequence of the past several consecutive days as the time-series input, and the second type is the static input vector of the encoded clinical baseline data. Categorical variables are encoded using one-hot encoding, and numerical variables are processed using mean normalization based on the range. The Long Short-Term Memory (LSTM) network updates the hidden state step by step with the temporal input. The output of the hidden state at the final time step is concatenated with the static input vector and then fed into a fully connected linear mapping layer. The Sigmoid activation function outputs the probabilities of acute exacerbation and disability, respectively, both of which are mapped to the interval between zero and one.

6. The adaptive early warning and intervention method for chronic obstructive pulmonary disease according to claim 1, characterized in that, The early warning judgment is executed collaboratively through real-time threshold judgment and adaptive rule base matching. This includes: comparing real-time blood oxygen saturation and respiratory rate values ​​with the baseline range of the corresponding activity state in the individual's health profile; generating a single indicator over-limit marker when the real-time blood oxygen saturation value is lower than the lower limit of the individual's baseline range and the duration exceeds a preset duration threshold, or when the real-time respiratory rate value is higher than the upper limit of the individual's baseline range and the duration exceeds a preset duration threshold; inputting the single indicator over-limit marker, risk probability, indicator change trend, and indicator composite relationship into the adaptive rule base for matching, and outputting an early warning level based on the matching result; the early warning level is divided into three levels: reminder level, intervention level, and emergency level.

7. The adaptive early warning and intervention method for chronic obstructive pulmonary disease according to claim 6, characterized in that, The trigger thresholds for each rule in the adaptive rule base are dynamically adjusted based on the patient's GOLD grade, history of acute exacerbations, and current risk probability of acute exacerbations. The adjustment method is as follows: starting from the baseline trigger threshold stored in the adaptive rule base, the baseline trigger threshold is adjusted by three factors: GOLD grade, number of acute exacerbations in the past year, and current risk probability of acute exacerbations. The three adjustment offsets are then added together to obtain the personalized trigger threshold applicable to the patient. The higher the GOLD rating, the more frequent the acute exacerbations, or the greater the probability of an acute exacerbation, the greater the magnitude of the adjustment offset that shifts the personalized trigger threshold in a more sensitive direction.

8. The adaptive early warning and intervention method for chronic obstructive pulmonary disease according to claim 1, characterized in that, The dynamic assembly of intervention content from the structured intervention knowledge base includes: filtering the scope of intervention content templates based on the warning level, where the intervention content corresponding to the reminder level warning is pushed to the patient end, the intervention content corresponding to the intervention level warning is pushed to both the patient end and the medical staff end, and the emergency level warning triggers the highest priority alarm and pushes the emergency medical reminder to the patient end and the emergency contact instruction to the medical staff end respectively; based on the medication plan, activity level and previous intervention response records in the individual health profile, the parameterized fields in the selected intervention content template are filled and adjusted to obtain the personalized intervention plan text; the personalized intervention plan text and the associated monitoring data summary are encapsulated into push messages to the patient end and push messages to the medical staff end respectively, and pushed synchronously to the corresponding terminals through the message channel.

9. The adaptive early warning and intervention method for chronic obstructive pulmonary disease according to claim 1, characterized in that, The process of updating the trigger thresholds and parameters of the dynamic risk prediction model based on feedback data includes: performing correlation analysis between patient symptom change data and subsequent physiological time-series monitoring data in the feedback data and the risk probability value and warning level at the time of warning triggering, and determining the validity label of this warning. The validity label includes valid warning and over-warning; for rules associated with over-warning, increasing the trigger threshold of the corresponding rule in the adaptive rule base or extending the duration requirement; for rules associated with missed reporting, decreasing the trigger threshold of the corresponding rule or shortening the duration requirement; supplementing the physiological time-series monitoring data and corresponding event labels in the feedback data into the training dataset of the dynamic risk prediction model, and incrementally training the dynamic risk prediction model with a preset retraining cycle to update the model's network parameters; wherein, the threshold change amount in each adjustment does not exceed a preset proportion of the current threshold.

10. An adaptive early warning and intervention system for chronic obstructive pulmonary disease (COPD), used to execute the adaptive early warning and intervention method for COPD as described in any one of claims 1 to 9, characterized in that, include: The data acquisition module is used to collect patients' real-time physiological signals and clinical baseline data, and integrate them to generate a multi-dimensional input dataset; The feature extraction module is used to preprocess and extract features from the physiological time-series data in the multi-dimensional input dataset to obtain risk feature vectors; The health profile module is used to build and dynamically update individual health profiles of patients based on multi-dimensional input datasets. The risk prediction module is used to input risk feature vectors and clinical baseline data into a dynamic risk prediction model and output the probability of acute exacerbation and the probability of disability. The early warning judgment module is used to perform early warning judgments and output the early warning level based on the probability of acute exacerbation, the probability of disability, individual health profile, real-time physiological signals and clinical baseline data, through real-time threshold judgment and adaptive rule base matching. The intervention plan generation module is used to dynamically assemble intervention content from a structured intervention knowledge base based on the warning level and individual health profile, generate personalized intervention plans, and push them to the patient and medical staff. The feedback update module is used to collect feedback data after an early warning, and update the trigger threshold of the adaptive rule base and the parameters of the dynamic risk prediction model based on the feedback data.

Images